Automatic Generation of Discharge Summary of EMRs Based on Multi-granularity Information Fusion

Discharge summaries are a significant component of electronic medical records, playing a crucial role in follow-up treatment and scientific research. However, there are few researches on automated discharge summary generation based on deep learning, and there is also a lack of available datasets. To address this, in this paper, we construct a small-scale dataset containing various types of entity information for the task of automated discharge summary generation from electronic medical records. In order to make full use of the rich entity information implied in medical records, we design a generation model based on a T5 architecture that encodes various types of entity information and incorporates the information contained in the entities into the encoder using multi-granularity fusion methods. Meanwhile, we use pointer-generator networks to enhance the model’s generalization capability. The experimental results show that the proposed dataset is challenging, and compared to the baseline models, the proposed model achieves significant improvements on the evaluation metrics. Additionally, ablation studies further demonstrate that incorporating entity information and pointer-generator networks positively contributes to the summarization quality of the model.